Zishuo Dong

Accurate identification of cracks at the pixel level on intricate asphalt pavements represents a crucial challenge in the domain of intelligent pavement assessment. The current advanced deep-learning networks encounter limitations in simultaneously capturing both the global context and local features of cracks, leading to discontinuous segmentation results and suboptimal recovery of local details. This paper proposes a robust architecture named Mix-Graph CrackNet to present an efficacious solution for this challenge. The Mix-Graph CrackNet, as proposed, is designed to mix the global context and local features multiple times, allowing for a comprehensively understanding of the essential features. Specifically, this paper develops the learnable parallel convolutional-Transformer mixing module to parallelly capture the sophisticated local features as well as the crucial global context. In addition, a new fusion unit is devised in the paper and deployed in the learnable parallel convolutional-Transformer mixing module. The proposed fusion unit is capable of effectively mixing contextual features extracted at both global and local scales while retaining an abundant level of textural details germane to the crack. Moreover, this paper constructs a graph-based skip connection that functions as a shortcut connecting the encoder and decoder, with the primary objective of mitigating information decay. The experimental results are remarkable, with the Mix-Graph CrackNet achieving F-measure and Intersection-Over-Union of 90.37% and 82.43%, respectively, on 1000 testing images. Based on the performance evaluations conducted on both public and private datasets, the proposed Mix-Graph CrackNet architecture demonstrates a significantly superior detection accuracy in comparison to several state-of-the-art models for semantic segmentation.

Pixel-level detection of expansion joints on complex pavements is significant for traffic safety and the structural integrity of highway bridges. This paper proposed an improved HRNet-OCR, named as expansion joints segmentation network (EJSNet), for automated pixel-level detection of the expansion joints on asphalt pavement. Different from the high-resolution network (HRNet), the proposed EJSNet modifies the residual structure of the first stage by conducting a Conv.+BN+ReLU (convolution+batch normalization+rectified linear unit) operation for each shortcut connection, which can avoid the network degradation. The feature selection module (FSM) and receptive field block (RFB) module are incorporated into the proposed EJSNet model to learn and extract the contexts at different resolution levels for enhanced latent representations. The convolutional block attention module (CBAM) is introduced to enhance the adaptive feature refinement of the network. Moreover, the shared multilayer perceptron (MLP) architecture of the channel attention module (CAM) is also modified in this paper. Experimental results demonstrate that the F-measure and intersection-over-union (IOU) attained by the proposed EJSNet model on 500 testing image sets are 95.14 and 0.9036, respectively. Compared with four state-of-the-art models for semantic segmentation (i.e., SegNet, DeepLabv3+, dual attention network (DANet), and HRNet-OCR), the proposed EJSNet model can yield higher detection accuracy on both private and public datasets.